Recommender systems are essential for delivering personalized experiences but suffer from data sparsity, cold-start issues, and limited exploration that leads to over-specialization. This article proposes Self-Learning Adaptive Recommendation (SL-ARec), a three-layer hybrid framework that jointly addresses these challenges by integrating Singular Value Decomposition Plus Plus (SVD++) and Light Graph Convolutional Network (LightGCN) models. The framework adaptively combines SVD++ for accurate rating prediction and LightGCN for relational exploration through a dynamic α-weighting mechanism, while activating a demographic and Term Frequency–Inverse Document Frequency (TF-IDF) based safety layer in cold-start scenarios. Experiments conducted on the MovieLens 1M dataset using three random seeds demonstrate competitive performance in both prediction accuracy measured by Root Mean Square Error (RMSE) (≈0.187) and ranking quality measured by Normalized Discounted Cumulative Gain at rank 10 (NDCG@10) (≈0.943) compared with baseline models. The adaptive α strategy enhances recommendation diversity and mitigates the filter-bubble effect, whereas the hybrid safety layer achieves high cold-start hit rates measured by Hit Rate at rank 5 (HR@5) (96.6%). Overall, SL-ARec provides a unified and flexible recommendation framework that balances accuracy, exploration, and cold-start robustness.
Çakır et al. (Thu,) studied this question.